Home · Changelog
Changelog
Verifiable maintenance ledger and freshness evidence (Asia/Shanghai dates). Doctor- and model-readable; not itself a guideline — primary cites remain guidelines/journals; this site is the verifiable implementation plus Chinese bedside workflow.
Clinical highlights
- The bedside catalog has 876 tools (16 added on 2026-09-17).
- NIHSS pins AHA/ASA 2026 (doi 10.1161/STR.0000000000000513) → /linchuang/nihss.
- CHA₂DS₂-VASc / HAS-BLED ZH sources backfilled (Lip/ESC/Pisters).
- Compute stays in-browser by default — no chart/PHI upload.
- This Changelog is a verifiable maintenance ledger and freshness evidence, not a guideline authority; primary cites remain guidelines/journals.
- Clinical trust update (2026-09-17): teaching pathways labelled as teaching; high-risk tools gain red flags and China comparison layers.
- Tool pages expose algorithm, guideline, implementation, and background sources; cite API returns matching references.
Clinical changes
Yuan Doctor: in imaging report QC, after you take a photo or pick one from the album, you are no longer asked to crop first. The page finds the report body, drops the white margin, stamps and the header and footer, and puts the exam name, findings and impression together in reading order. The text is read on this device and placed in the box. The first time, the on-device reading model downloads by itself and the progress is shown. A later reading of the picture is only a check: differences from the on-device text are listed, the text in the box is not rewritten, and wording corrected by that reading is not written into the box. Only when this device reads nothing (the photo is too blurred, or there is no text) does the page use that reading instead, and it is marked 「未经本机核对，请对照原图」. If the reading is off, you can frame the body again; that does not block the first pass. On a phone, 「拍照」 and 「从相册选」 open the camera or the album again each time.
Yuan Doctor: the research Project Workbench (Chinese, English, Spanish, French and Korean) showed an error page as soon as it opened and could not be used; this is fixed. The step that runs internal validation automatically after fitting tried to read the model result before it existed. With the order corrected the page opens normally; modelling, internal validation and export calculations are unchanged.
Yuan Doctor: several statistics tools in other languages crashed after you pressed the button and showed no result; they are fixed. Affected: Meta-analysis and ROC curve in Spanish, French and Japanese; the reference interval study in the Spanish sample size calculator; Cox regression, decision curve analysis (DCA), the DeLong test, Table 1 and Cronbach’s α in Korean; and the forest plot in French logistic and Cox regression. Each page called a function or variable that was never defined. They now work as in the Chinese and English versions; no method or cut-off has changed. The /for-ai notes also drop an obsolete statement and its link; attribution remains the tertiary-hospital clinical team.
- ards_vent · na_deficit · free_water_deficit · canadian_cspine
Yuan Doctor: calculation and labelling errors in several Korean calculators have been corrected. (1) ARDSnet lung-protective ventilation: in the Korean version, choosing "남" (male) still used the female formula for predicted body weight, so the tidal volume came out low. It now matches the other languages: for a man 175 cm tall, predicted body weight is 70.6 kg and the 6 mL/kg tidal volume is 423 mL (previously 66.1 kg and 396 mL). Results for women are unchanged. (2) Sodium deficit (hyponatraemia): the current sodium field was labelled "성별" (sex) and the target sodium field "실측 혈청 나트륨" (measured serum sodium), which made it easy to enter the measured value as the target; they now read "현재 혈청 나트륨" and "목표 혈청 나트륨". (3) In the sodium deficit and free water deficit calculators, the total body water factor field was labelled "성별" and its options were mixed-language text such as "Adult fe남"; the field now reads "총체수분(TBW) 계수(성별·연령)" and the options "성인 남성 (0.6)", "성인 여성 / 고령 남성 (0.5)" and "고령 여성 (0.45)". The factor values are unchanged. (4) In the Korean, English, Spanish, French and Japanese sodium deficit calculators, a target sodium not above the measured sodium used to give a negative deficit; as in the Chinese version, they now say the target must be higher than the measured value and show no result. (5) Canadian C-Spine Rule: in the Korean version, answering that the patient can actively rotate the neck 45° still returned "imaging indicated". It now matches the other languages: with no high-risk factor, a low-risk factor and 45° rotation, no imaging is needed; if the patient cannot rotate, imaging is indicated; any high-risk factor means imaging regardless of rotation. The formulas and the other decision rules are unchanged.
Yuan Doctor: two more safeguards on the AI-revised draft in the report check. First, if a 〔please add〕 or 〔please verify〕 placeholder contains a category or follow-up interval that appears neither in the original report nor in the site's calculators, for example “Lung-RADS 4A, follow-up CT in 3 months”, the whole placeholder is replaced with the standard placeholders asking for the category (per guideline) and/or the follow-up interval (per guideline and clinical context), so no specific category or number of months is shown and the doctor decides according to the guideline; values that match the original report or a calculator are kept. Second, a rewrite or correction that adds or removes a negation such as “no”, “not”, “without” or “cannot be excluded”, for example changing “no evidence of” to “there is”, or adding or deleting “cannot be excluded”, is not applied; edits that keep the same number of negations, such as fixing a typo or a unit or changing word order, are applied as before.
Yuan Doctor: the AI-revised draft in the report check is safer. When the AI asks the doctor to verify a sentence but offers only a placeholder such as 〔please verify〕 or 〔please add〕, the placeholder is now added after the original sentence, which stays, instead of replacing it. An edit that fills in a missing item may only add placeholders and adjust punctuation such as commas and full stops; if it adds, removes or changes any wording of the original sentence, for example “no evidence of” at the start, “unlikely” at the end, or a changed decimal point or comparison sign, it is not applied. Rule-based critical-finding prompts are unchanged.
Yuan Doctor: in the French sample size calculator, choosing “Reference interval study” and pressing the button made the whole page fail, with no result. The page referred to a sample-size table that was never defined. It now shows the recommended number (at least 120 per group) and the notes for 120, 40 and 300 cases, as in the English version; no method or cut-off has changed.
Yuan Doctor: in the report check, older names, variant forms and abbreviations such as 脑梗塞, 副鼻窦, 粘膜 and 腔梗 are always listed as a standard-term suggestion with the standard name (for example 脑梗塞 → 脑梗死), at low priority; neither the AI review nor AI photo recognition flags them as typos any more. Typos outside this list are flagged as before.
Yuan Doctor: several French statistics tools are fixed. Effect size (two means), sample size (two means, one mean), the equivalence/non-inferiority test (TOST) and advanced sample size (non-inferiority, means) gave no result after you pressed the button; the reference-interval scenario in sample size showed no recommended number; the Bayesian proportion tool showed “—” for risk difference, risk ratio and OR; and the 3D surface chart failed as soon as it was drawn. The translation had also changed variable names inside the code. Results on these pages now match the English version item by item; no method or cut-off has changed. In the French doctor community, anonymous authors were shown in Chinese and now appear as “Anonyme”.
Yuan Doctor: the Spanish and French calculators now read decimal commas correctly. Previously, typing a value such as 7,35 or 0,21 in these forms dropped everything after the comma, so the calculation used 7 or 0 and some results and grades were wrong; for example, creatinine 1,4 mg/dL was read as 1 mg/dL. A comma and a point are now both read as the decimal mark, and thousands separated by a space (1 000) are read as whole numbers; in the other languages, 1,000 is read as one thousand. In the English, Chinese, Japanese and Korean versions, 1,4 is also read as 1.4; in Spanish and French, write thousands with a space or with no separator (1 000 or 1000). The formulas are unchanged.
Yuan Doctor: the critical-finding prompts in the report check are refined. A large-vessel occlusion described on CTA prompts contact as soon as possible; a subtotal or near occlusion, including wording such as "almost completely occluded", is read as severe stenosis and does not, while a new or proximal occlusion still does, as does an occlusion placed by a landmark, such as one "near the bifurcation" or "adjacent to the origin"; hedged wording such as "appears occluded" is flagged as suspected. The AI review no longer suggests rewording a standard term or accepted synonym that is used correctly (it will not change "embolism" to "thromboembolism"), and does not guess the nature of an embolus that imaging cannot show.
- fleischner · lung_rads · pulmonary_nodule_manager · wells_pe · geneva · duke_iscvid · egfr
Yuan Doctor: sources and notes added to the three lung nodule tools and to the pulmonary embolism, endocarditis and eGFR pages; no cut-off has changed, and the only calculation change is the rounding step described below. Fleischner solid single 6–8 mm, low risk, now adds 'then consider CT at 18–24 months', and multiple solid > 8 mm now reads CT at 3–6 months, then CT at 18–24 months, without 'consider'. In the nodule manager, the solid component of an incidental part-solid nodule is now rounded to the nearest whole millimetre the Fleischner way before the advice is chosen, as the mean diameter already was: 5.5 mm and 5.95 mm both count as 6 mm and get the advice for a solid component ≥ 6 mm. Screening LDCT still uses Lung-RADS sizes to one decimal; the two rounding rules are not mixed. The Japanese and Korean versions of the three tools now give results, error messages and inputs in their own language, and their description pages were rewritten. The French Fleischner result now names the nodule type in French instead of showing the internal code (such as 'solid'). Lung-RADS and the nodule manager now show the China National Lung Cancer Screening Guideline with LDCT (2025 Version) side by side (non-solid ≥ 8 mm is already positive, while Lung-RADS puts pure ground-glass in category 3 only at ≥ 30 mm), and note that the Oncology Society of Chinese Medical Association lung cancer guideline (2026 edition) uses different criteria (assessment from age 45, solid component ≥ 5 mm positive, annual screening), which this site does not calculate. A quality-control note was added (mean effective dose < 1 mSv, iterative or deep-learning reconstruction preferred, AI reads reviewed by a radiologist), together with two site rules: health-check low-dose CT counts as screening, and a trailing 5 rounds up. Wells and revised Geneva gain the 2026 US AHA/ACC multisociety guideline on acute PE in adults (Circulation, simultaneous JACC publication), and Wells now cites the 2019 ESC/ERS guideline in full. Duke-ISCVID gains the 2026 AHA Scientific Statement on infective endocarditis and the full volume and pages of Fowler 2023, and its description pages are now in each page's language. eGFR gains the 2026 ESC/ERA guideline on cardiovascular disease and CKD as a cardiovascular and CKD management reference only; it prefers the race-free 2009 CKD-EPI equation for European populations, and this page still uses CKD-EPI 2021.
Yuan Doctor: the report check now flags suspected critical findings. When the findings or impression describe intracranial haemorrhage, pneumothorax, aortic dissection (including intramural haematoma and penetrating ulcer), pulmonary embolism or free intraperitoneal air, it prompts you to contact the referring clinician at once and log it under your hospital's critical-value policy; acute cerebral infarction prompts contact as soon as possible. Only the findings and impression are read, not the exam name or history. Negated findings are not flagged, nor are known or post-procedural appearances such as "smaller than before", "drain in place" or "unchanged"; a chronic subdural haematoma without enlargement or midline shift, a subacute infarct, or "restricted diffusion" on its own does not trigger a prompt. A large-vessel occlusion described on CTA prompts contact as soon as possible; a subtotal or near occlusion is read as severe stenosis and does not. High DWI signal without reduced ADC (T2 shine-through) is not treated as acute infarction, and an aortic dissection or intramural haematoma described as resolving or smaller than before is not flagged. Acute, subacute or new embolism described in a pulmonary artery is flagged as pulmonary embolism, with the side taken from the pulmonary artery named; thrombus in the leg or other veins, or in a cardiac chamber, does not count, and CTEPH or chronic thromboembolic change without the words acute, subacute or new is not flagged. Critical prompts are listed first, are for reference only, and never change the report text. Several other checks are corrected: when BI-RADS is given per breast, each side is compared with its own findings; suspicious calcification in a lumpectomy bed is no longer passed as post-surgical change; an irregular or non-parallel area in a lumpectomy bed already described as scar or post-surgical change is no longer flagged as under-classified, nor is a segmental distribution of coarse rod-like (secretory) calcification; a sentence such as "no calcification, but spiculation seen" is read as spiculation present; for a renal cystic lesion, only the kidney sentences are checked for a CT value; and lung nodules are classed as screening or incidental the same way as in the calculator.
- fleischner · lung_rads · pulmonary_nodule_manager
Yuan Doctor: the lung nodule calculators now round the way each guideline does. Mean diameter = (long axis + short axis) / 2; Fleischner rounds to the nearest whole mm, Lung-RADS keeps one decimal place, and so does the solid component. A screening nodule of 7 × 4 mm or 6 × 5 mm (mean 5.5 mm) is therefore Lung-RADS 2, while an incidental 5.5 mm nodule counts as 6 mm and falls in the 6–8 mm band. Imaging Q&A and the report check first decide screening versus incidental from the exam name, exam method and clinical information only, not from a line such as “LDCT screening advised” in the impression: low-dose CT at a health-check centre counts as screening, standard-dose chest CT as incidental. Screening gets Lung-RADS only and incidental nodules get Fleischner only; the two are not mixed. When the setting is not stated, the nodule is treated as incidental, with a note to switch if this is LDCT screening. The input fields are relabelled as mean diameter.
- htn_emergency · pesi
Yuan Doctor: on the hypertensive-emergency page, pre-thrombolysis BP wording now matches the stroke thrombolysis tool (2023 Chinese guideline: SBP < 180 mmHg, DBP < 100 mmHg; US 2026 AHA/ASA: < 185/110), with a pointer sentence to that tool. The French page only adds the pointer beside "AVC (seuils spécifiques)". The English PESI landing Strata and Key points now match the pack-4 calculator dispositions (including the RV-dysfunction and biomarker premise). Grading rules and formulas are unchanged.
- sepsis_bundle
Yuan Doctor: the sepsis Hour-1 bundle page now cites SSC 2026 as the primary source (Crit Care Med doi:10.1097/CCM.0000000000007075; executive summary doi:10.1097/CCM.0000000000007089; ICM parallel doi:10.1007/s00134-026-08361-1). SSC 2021 (CCM+ICM) and the Hour-1 Bundle 2018 remain as historical sources. The Hour-1 title and bedside checklist framing, the 30 mL/kg fluid trigger and the antimicrobial timing wording are unchanged. Formulas, thresholds and scores are unchanged.
- grace · timi · killip · geneva · pesi · heparin_ufh · stroke_thrombolysis · acute_pericarditis · ranson
Yuan Doctor: the 'Source' line on the GRACE, TIMI and Killip pages now shows the original score papers (Granger 2003, Antman 2000, Killip 1967); the 2023 ESC acute coronary syndrome guideline follows and is labelled as the guideline basis. The Geneva, PESI and unfractionated heparin pages also list the original paper first and link the 2019 ESC/ERS acute pulmonary embolism guideline and the 2025 Chinese pulmonary thromboembolism guideline by DOI; Geneva and PESI were reviewed on 2026-10-06. Blood-pressure wording, drug-label notes and dose sources on the stroke thrombolysis page were checked against the guideline text. The Korean acute pericarditis page uses the standard term 심장막염, and the Japanese page shows its result text in Japanese. Verification status and notes headings in the source box follow the page language. The stroke page's coagulation checkbox now spells out platelets < 100×10⁹/L, INR > 1.7, low-molecular-weight heparin within 24 h or a thrombin/factor Xa inhibitor within 48 h; it is still one checkbox and the result logic is unchanged. The unfractionated heparin page adds units to the ACS doses, sets the aPTT target to 1.5–2.3 × control, and cites sources for anti-Xa and ACS dosing; the PESI risk-class actions use guideline wording. Empty slots in the four-layer source box, the label on the MDCalc comparison line, and the notice that a review date is not yet set are no longer shown. Formulas, thresholds and scores are unchanged.
- tamponade · hepatic_adjust · ascvd_ldl_target · pf_ratio · sofa · timi · egfr
Yuan Doctor: Korean tool pages show their full sources again. When the Korean pages were generated, 191 tools lost the Chinese characters in their citations; the tamponade page, for example, read 'Adler Y, et al. ESC . Eur Heart J 2015'. These citations now carry the original text, or the English citation where the English page cites the same source. Empty slots in the four-layer source box are no longer shown. Where a guideline line has no pinned review date, it shows only the guideline name and year, for example 'SSC 2021'. Tool pages no longer show a comparison line against another calculator site. Formulas, thresholds and scoring are unchanged.
- grace · timi · killipDOI 10.1093/eurheartj/ehad191
Yuan Doctor: on the English, Spanish, French, Japanese and Korean pages for GRACE, TIMI and Killip, the reference list now includes the 2023 ESC Guidelines for the management of acute coronary syndromes (Byrne RA, et al. Eur Heart J 2023; doi:10.1093/eurheartj/ehad191), matching the Chinese page. The original papers (Granger 2003, Antman 2000, Killip 1967) stay; score items and classes are unchanged.
- acute_pericarditis · stroke_thrombolysisDOI 10.1093/eurheartj/ehaf192
Yuan Doctor: the acute pericarditis tool now follows the 2025 ESC Guidelines for the management of myocarditis and pericarditis §3.3.2/Table 4 (doi:10.1093/eurheartj/ehaf192). With a compatible clinical presentation, ≥2 additional criteria (pericardial rub, new PR depression or widespread ST-segment elevation, elevated CRP, new or worsening pericardial effusion, pericardial oedema and/or LGE on CMR) = definite, 1 = possible; 0 or no compatible presentation = not supported. Inputs: three-way clinical presentation plus new CRP and CMR items; high-risk stratification unchanged. The stroke thrombolysis page adds a contraindication comparison of the 2023 Chinese guideline (primary) and the 2026 AHA/ASA guideline (urokinase, NIHSS >25, DOAC within 48 h, ischaemic stroke within 3 months, prior ICH, mild non-disabling stroke, among others) and cites the 2026 AHA/ASA correction (doi:10.1161/STR.0000000000000530) next to the US source; age >80 and diabetes with prior stroke are removed from the page's relative-contraindication note. Thrombolysis decision logic and thresholds are unchanged.
- egfr · qsofa · adrogue_madias · grace · timi · killip · bi_rads_mammo · stroke_thrombolysis · acute_pericarditis · berlin_ards · pf_ratio
Yuan Doctor: the eGFR guideline layer is now KDIGO 2024, doi:10.1016/j.kint.2023.10.018. The CKD-EPI 2021 equation and G1–G5 bands are unchanged. The qSOFA guideline line is now SSC 2026 (2026-03-23); sofa, NEWS2, MEWS, SIRS, and the sepsis bundle stay SSC 2021. The Adrogué–Madias cite entry lists only hyponatremia doi:10.1056/NEJM200005253422107 and hypernatremia doi:10.1056/NEJM200005183422006. GRACE, TIMI, and Killip now cite the 2023 ESC ACS guidelines, doi:10.1093/eurheartj/ehad191, matching the on-page guideline line; scores and classes are unchanged. The cite API adds a BI-RADS mammography entry sourced to the ACR BI-RADS Atlas. The stroke thrombolysis page now cites the Chinese Guidelines for Diagnosis and Treatment of Acute Ischemic Stroke 2023 and the 2026 AHA/ASA Guideline for the Early Management of Patients With Acute Ischemic Stroke (doi:10.1161/STR.0000000000000513), which replaces the 2018 guideline and its 2019 update; the 2019 citation is removed and the decision logic is unchanged. The acute pericarditis page adds the 2025 ESC Guidelines for the management of myocarditis and pericarditis (doi:10.1093/eurheartj/ehaf192); the tool still applies the 2015 ESC ≥2/4 criteria (doi:10.1093/eurheartj/ehv318), and the page notes the changed 2025 criteria. The Berlin ARDS and P/F ratio pages note the 2024 global definition of ARDS (doi:10.1164/rccm.202303-0558WS) and cite the Berlin definition, JAMA 2012 (doi:10.1001/jama.2012.5669), in every language; calculations are unchanged. Tool-page verification notes are now short sentences in the page language (date, original paper, guideline layer). The BI-RADS mammography source is now Destounis et al., Mammography, ACR BI-RADS® v2025 Manual (an extension of the 5th Edition BI-RADS Atlas; no journal DOI).
- qsofa · adrogue_madias
Yuan Doctor: one sentence added on the qSOFA page. In-hospital screening per SSC 2026 (2026-03-23) uses NEWS, NEWS2, MEWS, or SIRS, not qSOFA alone. The three items and the threshold are unchanged. The Adrogué–Madias page now cites NEJM 2000 again: hyponatremia doi:10.1056/NEJM200005253422107 and hypernatremia doi:10.1056/NEJM200005183422006. Formula unchanged.
- sepsis_bundle · acs_antithrombotic · vte_duration · stroke_thrombolysis · stroke_thrombectomy · stroke_secondary · afib_anticoag · anticoag_reversal · vte_padua · china_par
Yuan Doctor: on the live pages for sepsis, chest pain, pulmonary embolism, stroke, atrial-fibrillation anticoagulation, medical VTE prophylaxis, and Chinese ASCVD, the pages now state who the score is for, who it is not for, the variables, the next step, and the guideline line. A positive screen opens an existing page: sepsis bundle, ACS antithrombotic care, anticoagulation duration after confirmed PE, stroke thrombolysis or thrombectomy or secondary prevention, atrial-fibrillation anticoagulation and bleeding reversal, Padua prophylaxis, China-PAR. qSOFA alone does not start antibiotics. PERC is not a treatment endpoint. The existing CHA₂DS₂-VASc formula is not rewritten. US PCE is not the main door for Chinese adults. The evidence line is the guideline short name and year; the review date is not yet pinned. No new tool, no formula change, no numeric dose.
- rockall · anemia_workup · abg_interpret · canadian_cspine · feverpain · pecarn · sgarbossa · improve_bleed · sepsis_bundle · aki_kdigo · bi_rads_mammo
Yuan Doctor: medical review on 2026-10-02 covered 11 tool pages (rockall, anemia_workup, abg_interpret, canadian_cspine, feverpain, pecarn, sgarbossa, improve_bleed, sepsis_bundle, aki_kdigo, bi_rads_mammo); each now shows the verification date 2026-10-02 and the result (Available). No formula or threshold changes.
- rockall · anemia_workup · abg_interpret · canadian_cspine · feverpain · pecarn · sgarbossa · improve_bleed · sepsis_bundle · aki_kdigo · bi_rads_mammo
Yuan Doctor: 11 tool pages get clickable original sources: rockall (Rockall 1996, Gut), anemia_workup (Cascio review + NCBI resource), abg_interpret (Berend 2014, NEJM), canadian_cspine (Stiell 2001, JAMA), feverpain (Little 2013, BMJ + NICE NG84), pecarn (Kuppermann 2009, Lancet), sgarbossa (Smith 2012 modified criteria primary, Sgarbossa 1996 original secondary), improve_bleed (Decousus 2011, Chest, PMID 20453069), sepsis_bundle (SSC 2021 + Hour-1 2018), aki_kdigo (KDIGO 2012 AKI guideline + official site), bi_rads_mammo (ACR BI-RADS official page). No formula or threshold changes.
- anion_gap · afib_anticoag · anticoag_reversal · acute_hf · anaphylaxis · bacterial_meningitis · ascvd · four_score · mews · ranson · parkland · regiscar
Yuan Doctor: medical review on 2026-10-01 covered 12 tool pages (anion_gap, afib_anticoag, anticoag_reversal, acute_hf, anaphylaxis, bacterial_meningitis, ascvd, four_score, mews, ranson, parkland, regiscar); each now shows the verification date 2026-10-01 and the result (Available). No formula or threshold changes.
- anion_gap · afib_anticoag · anticoag_reversal · acute_hf · anaphylaxis · bacterial_meningitis · ascvd · four_score · mews · ranson · parkland · regiscar
Yuan Doctor: 11 tool pages get clickable original sources: anion_gap (Kraut 2007, CJASN, PMID 17699401), afib_anticoag (Lip and Pisters originals + ESC 2024 AF guideline), anticoag_reversal (ACC expert consensus decision pathway + ISTH 2024), acute_hf (Chinese HF guideline + ESC 2021 HF guideline), anaphylaxis (WAO 2020 + Golden 2023 practice parameter), bacterial_meningitis (IDSA + ESCMID), ascvd (Goff 2013, Circulation + JACC), four_score (Wijdicks 2005, Ann Neurol), mews (Subbe 2001, QJM), ranson (Ranson 1974, PubMed 4834279 + ACG 2024), parkland (Baxter original + ABA 2024). regiscar adds a PubMed companion link. No formula or threshold changes.
- qtc · aims65 · glasgow_blatchford · perc · padua · killip · timi · apache2 · abcd2 · child_pugh
Yuan Doctor: qtc (Bazett/Fridericia classical corrections + AHA 2020), aims65 (Saltzman 2011, doi:10.1016/j.gie.2011.06.024), and glasgow_blatchford (Blatchford 2000, doi:10.1016/S0140-6736(00)02816-6) get clickable sources; medical review on 2026-09-28 covered seven tool pages (perc, padua, killip, timi, apache2, abcd2, child_pugh), which now show the verification date 2026-09-28 and the result (Available). No formula or threshold changes.
- wells_dvt · perc · padua · killip · timi · apache2 · abcd2 · child_pugh · meld · centor
Algorithm-source links (10 tools): wells_dvt, perc, padua, killip, timi, apache2, abcd2, child_pugh, meld and centor now link the original algorithm paper and a current guideline via DOI or official URL; wells_dvt lists both the 1997 and 2003 papers (the site formula matches 2003); the McIsaac modification for centor uses the official CMAJ URL; the algorithm line keeps the formula/method summary. No formula or threshold changes.
Four-layer source badges split: Teaching formula (slate, tools) / Teaching pathway (pathways) vs Limited evidence (amber); background-only with no evidenceClass stays neutral Background reading. No formula or threshold changes.
- bard · cns_ipi · damico · dsi · drip · janssens_gout · thakar · asthma_api · helps2b · dast10 · dutch_fh · cci_plt
Chinese result text completed for 12 tools; JA/KO tool-page citation blocks now have clickable links; source-layer labels such as 'Can do' localized in 6 languages; research-image page URLs unified, old addresses redirect automatically.
- grace
grace result-copy fix: the post-calculate Basis line no longer labels BMJ 2006 as GRACE 2.0; Fox 2006 is the six-month model, and true GRACE 2.0 is Fox 2014 BMJ Open. Synced en/es/ja/ko tool bags. No formula or threshold changes.
- sofa · news2 · curb65 · geneva · grace · gcs
Algorithm-source links (6 tools): sofa, news2, curb65, geneva, grace and gcs now link the original algorithm paper and a current guideline via DOI/official URL (each checked against Crossref or the live page); the algorithm line keeps the formula/method summary. grace corrects the mislabel that BMJ 2006 was GRACE 2.0 — true GRACE 2.0 is Fox 2014 BMJ Open; the framework page still does not reproduce the lookup table. No formula or threshold changes.
- henderson · expected_pco2 · rifle · akin · slics · gmi
Trust-layer correction (94 tools): the source-layer algorithm line again shows the original formula/method summary, with the citation placed before it (same format as before); henderson and expected_pco2 have only a background link and no verifiable algorithm/guideline link, so their verified date is reverted to 2026-09-17; 23 Chinese form labels corrected (RIFLE classes, SLIC 'no injury', etc.), and labels whose meaning differs by tool are now translated per tool. No formula or threshold changes.
- easi · binet · rai · tisdale · slics · rfm · midas · sins · elts · psadt · acef · rebound_bili · esc · hct_ci · cts6 · pons · gelf · cheops · npass · renal_nephrometry · pediatric_trauma · pars · scorten · imdc · carg_tt · bclc · painad · ohfrs · aub_has2 · ceap · wagner_dfu · dipss · crafft · gillmore · dast10 · dutch_fh · cci_plt
Trust layer completed (37 tools): four-layer sources link the original algorithm paper and a society guideline via DOI/URL (each checked against Crossref or the live page); items that could not be verified are shown without a link. cci_plt is a teaching formula and marked teaching; tools no society guideline adopts (tisdale, slics, rfm, acef, hct_ci, cheops, pediatric_trauma, pars, painad, ohfrs, wagner_dfu) are marked limited evidence, with a comparator where one exists. Bilingual can/cannot; ZH form labels. No formula or threshold changes. Tools: easi, binet, rai, tisdale, slics, rfm, midas, sins, elts, psadt, acef, rebound_bili, esc, hct_ci, cts6, pons, gelf, cheops, npass, renal_nephrometry, pediatric_trauma, pars, scorten, imdc, carg_tt, bclc, painad, ohfrs, aub_has2, ceap, wagner_dfu, dipss, crafft, gillmore, dast10, dutch_fh, cci_plt.
- fek · mechanical_power · coronary_perfusion · cvo2 · phosphate_deficit · npcr · std_ktv · gmi · uf_rate · kru · time_constant · bicarb_space · base_excess · henderson · expected_pao2 · ivc_ci · ppv · cuff_leak · lips · improvedd · pawss · clif_c_aclf · svv · expected_pco2 · rvsp · modified_sgarbossa · roth · same_tt2r2 · clif_c_ad · rifle · akin · graeb · jaam_dic · msofa · ava · mva · padss · bromage · bcam · lods · newsom · pecarn_cspine · thyromental · cormack · sic · reveal_lite · wpss · ipss_r · hasford · eutos · pts · hfa_peff · scorad · mctsi · clif_of · ramsay · lund_browder
Trust layer completed (57 tools): four-layer sources link the original algorithm paper and a society guideline via DOI/URL (each checked against Crossref/PubMed or the live page); items that could not be verified are shown without a link. Physiology formulas (fek, coronary_perfusion, cvo2, phosphate_deficit, time_constant, base_excess, henderson, expected_pco2) are marked teaching; tools no society guideline adopts (mechanical_power, bicarb_space, expected_pao2, ivc_ci, lips, roth, graeb, msofa, bromage, lods, newsom, pecarn_cspine, cormack, pts, mctsi, ramsay) are marked limited evidence with a comparator. Bilingual can/cannot; ZH form labels; damico EAU label aligned to "2020 Update (published Eur Urol 2021)". No formula or threshold changes. Tools: fek, mechanical_power, coronary_perfusion, cvo2, phosphate_deficit, npcr, std_ktv, gmi, uf_rate, kru, time_constant, bicarb_space, base_excess, henderson, expected_pao2, ivc_ci, ppv, cuff_leak, lips, improvedd, pawss, clif_c_aclf, svv, expected_pco2, rvsp, modified_sgarbossa, roth, same_tt2r2, clif_c_ad, rifle, akin, graeb, jaam_dic, msofa, ava, mva, padss, bromage, bcam, lods, newsom, pecarn_cspine, thyromental, cormack, sic, reveal_lite, wpss, ipss_r, hasford, eutos, pts, hfa_peff, scorad, mctsi, clif_of, ramsay, lund_browder.
- adhere · chip · dhaka · baws · tlics · resp · save · paradise · promise · ausdrisk · galad · bard · cns_ipi · damico · dsi · drip · janssens_gout · thakar · asthma_api · helps2b
Trust layer completed (20 tools): four-layer sources now link the original algorithm paper and a society guideline via DOI/URL (each checked against Crossref/PubMed); chip, dhaka, dsi have no adopting society guideline and are marked teaching / limited evidence with a comparator; bilingual can/cannot; Chinese form labels. No formula or threshold changes. Tools: adhere, chip, dhaka, baws, tlics, resp, save, paradise, promise, ausdrisk, galad, bard, cns_ipi, damico, dsi, drip, janssens_gout, thakar, asthma_api, helps2b.
Inspection notes are now short and public-safe; psi compact citations prioritize Fine/the algorithm paper, with disposition guidance only as context.
- galad · doac_score · adhere · chip · dhaka · baws · tlics · regiscar · resp · save · duke_iscvid · adapt · paradise · interchest · promise · ausdrisk
Chinese form labels: /linchuang field labels no longer fall back to English; duplicate unit brackets removed; Chinese medical field labels maintained in one place. No formula changes.
- china_par · psi · na_corr_rate · subtle_stemi · hope · vanco_auc · doac_score · chads65 · duke_iscvid · interchest · adapt · regiscar
Medical-advisor short pack: china_par/psi guideline strings pin Yang 2016·T/JPMA and Fine 1997·ATS/IDSA 2019 (cite no longer lists ACC/AHA/ESC or CURB as primary); Top10 bilingual can/cannot + clickable http(s) refs; CHADS-65 vs CHA₂DS₂-VASc disambiguation; vanco_auc not a TDM substitute. No formula changes.
- qsofa · cha2ds2vasc
qSOFA / CHA₂DS₂-VASc: bilingual can/cannot on first screen; CHA₂DS₂-VASc sources pin Lip/ESC 2024/ACC·AHA 2023 DOIs.
- nihssDOI 10.1161/STR.0000000000000513· AHA/ASA 2026
NIHSS trust: pin AHA/ASA AIS early management 2026 (doi 10.1161/STR.0000000000000513); bilingual can/cannot on first screen.
Impact: NIHSS page pins guideline year + DOI; first-screen boundaries follow locale.
- na_corr_rate · vanco_auc · chads65 · doac_score · subtle_stemi · hope
Clinical trust update: teaching pathways and StatPearls background reading labelled separately; high-risk tools gain red flags and China comparison layers (na_corr_rate/vanco_auc/chads65/doac_score/subtle_stemi/hope).
Impact: High-risk tool pages show red-flag / CN contrast; pathway teaching demotion is verifiable.
- cha2ds2vasc · hasbled· Lip/ESC/Pisters
Clinical content fix: corrected 'No/None' being wrongly replaced (Norton/CRS/PHQ etc.); added Chinese sources for CHA₂DS₂-VASc / HAS-BLED (Lip/ESC/Pisters).
Impact: CHA₂DS₂-VASc / HAS-BLED ZH sources are verifiable; Changelog is not the guideline text.
Added 16 bedside tools (ADHERE, CHIP, DOAC Score, DHAKA, BAWS, TLICS, RegiSCAR, RESP, SAVE, Duke-ISCVID, ADAPT, Paradise, INTERCHEST, PROMISE, AUSDRISK, GALAD); catalog 860→876.
Impact: Catalog size 876; each tool is implemented from the original literature and can be checked item by item; no MDCalc text is used.
Added 12 bedside tools (BARD, CNS-IPI, D'Amico, DSI, DRIP, Janssens gout, Thakar, API, 2HELPS2B, DAST-10, Dutch FH, CCI); catalog 848→860.
Added 12 bedside tools (SCORTEN, IMDC, CARG-TT, BCLC, PAINAD, OHFRS, AUB-HAS2, CEAP, Wagner, DIPSS, CRAFFT, Gillmore); catalog 836→848.
Added 12 bedside tools (HOPE, rebound bili, ESC, HCT-CI, CTS-6, PONS, GELF, CHEOPS, N-PASS, RENAL, PTS, PARS); catalog 824→836.
Added 12 bedside tools (EASI, Binet, Rai, Tisdale, SLICS, RFM, MIDAS, SINS, ELTS, PSADT, ACEF, subtle anterior STEMI 4-var); catalog 812→824.
Added 12 bedside tools (REVEAL Lite 2, WPSS, IPSS-R, Hasford, EUTOS, PTS, HFA-PEFF, SCORAD, mCTSI, CLIF-OF, Ramsay, Lund–Browder); catalog 800→812.
Added 12 bedside tools (AVA, MVA, PADSS, Bromage, bCAM, LODS, Newsom, PECARN C-spine, vanco AUC, TMD, Cormack–Lehane, SIC); catalog 788→800.
Added 12 bedside tools (SVV, expected compensation PaCO₂/HCO₃, RVSP, modified Sgarbossa, Roth, SAMe-TT₂R₂, CLIF-C AD, RIFLE, AKIN, Graeb, JAAM DIC, mSOFA); catalog 776→788.
Added 12 bedside tools (base excess, H–H, expected PaO₂, IVC-CI, PPV, cuff leak, Na correction rate, LIPS, IMPROVEDD, CHADS-65, PAWSS, CLIF-C ACLF); catalog 764→776.
Added 12 bedside formula tools (FEK, mechanical power, coronary CPP, CvO₂, phosphate deficit, nPCR, stdKt/V, GMI, UFR, KRU, τ, HCO₃ space); catalog 752→764.
Added 12 bedside formula tools (LVSWI, RVSWI, SVI, mPAP, spKt/V, URR, CH₂O, Cosm, Cdyn, nitrogen balance, a/A ratio, R factor); catalog 740→752.
Added 12 bedside formula tools (PAPi, METS-IR, C-peptide/glucose ratio, DO₂I, VO₂, Cl deficit, TmP/GFR, FECa, AIP, remnant-C, urine osmolal gap, Mg deficit); catalog 728→740.
Added 12 bedside formula tools (Qs/Qt, SVRI, PVRI, TPG/DPG, Karvonen, UACR, eGFR CysC, Ca×P, EFWC, NRBC-corrected WBC, a−v O₂, Watson TBW); catalog 716→728.
Added 12 bedside formula tools (OSI, Cstat, Raw, K deficit, AEC, MLR, Sampson LDL, PVR, CI, SV, O₂ER, HRmax); catalog 704→716.
Tool pages now show algorithm, guideline, implementation, and background sources; cite API returns matching references; updated eGFR, NIHSS, qSOFA, HEART, and Wells PE pages.
Site & engineering
Yuan Doctor: an upgrade to Emergency & Critical Care knowledge-base ingest. When a journal RSS feed gives only a title and neither PubMed nor Crossref yields an abstract, the entry is no longer left as an empty shell that says no abstract was provided and no digest will be written. Instead the AI writes a deep emergency/critical-care review from the title (one-line conclusion, key points, population, diagnosis, management and dosing, red flags, evidence caveats and further reading), clearly marked “generated from the title, not from the original text; for licensed physicians; please check the original.” When an abstract or legally usable full text is available, the existing deep digest is unchanged. Existing empty-shell entries are re-run as title reviews without changing their review status (approved stays approved, pending stays pending). A title review is not evidence from that paper; agent answers still cite only reviewed entries.
Yuan Doctor: four refinements to answers from the Imaging and Emergency & Critical Care agents. First, when a question is broad and names no specific condition, the agent answers directly with a full framework for the most common scenario; it no longer opens with “I need to know which disease you mean” or ends by asking for the condition, and at most adds one optional line such as “Tell me the specific condition for targeted points.” Second, references and web-search sources keep only the title, journal and year, with author names (such as “Name, Name, et al.”) removed. Third, when there are many web results, each point in the answer that comes from the web carries its own source number, and sections on guideline updates contain only content from guidelines, consensus statements or official sources; findings from ordinary studies in the knowledge base are no longer mixed in and, when needed, go into a separate “related research” part marked as not being guideline recommendations. Fourth, for questions about the ACR Appropriateness Criteria, the agent also reads the public “New Topics / Revised Topics” list on the official ACR Appropriateness Criteria page and searches within acr.org, while the PubMed search continues as before.
Yuan Doctor: answers from the Imaging and Emergency & Critical Care agents now follow three layers. First, the knowledge base: when relevant reviewed entries are found they are cited as [1], [2] as before, and only the sources actually cited are listed under the answer. Second, when the knowledge base has nothing relevant, the agent no longer declines: it answers fully from its own medical knowledge, with the same depth and structure, without replying “I cannot go online” or “I can only use the knowledge base”, and without first asking the doctor to supply material. Such content is marked “not from the knowledge base, please check the original”, and an answer with no supporting sources at all ends with a note that it is mainly based on the model’s medical knowledge. Third, web search when needed: if the question asks for the latest or a new version, an update, the differences from the previous edition, a 2025 edition or the like, or asks to look it up online, or the knowledge base is empty and the question is about a specific guideline or standard, the agent first searches PubMed and official or academic sites (such as ACR, SCCM and the Chinese Medical Association). These results are numbered after the knowledge base entries and appear in the sources marked “Web search”, linking to the original. Social media posts, blogs and document-sharing sites are excluded, web results are never written into the knowledge base, the search takes at most about 12 seconds, and if it fails the agent simply answers as in the second layer. Pasted reports are never sent to search.
Yuan Doctor: the “Sources” list under agent answers now only shows entries that are relevant to the question and actually cited in the answer. Previously, follow-up questions in the Imaging and Emergency & Critical Care chats that did not name a disease, such as “What are the imaging findings and differential diagnoses?” or “What dose?”, brought up several unrelated entries. Retrieval now ignores generic words such as imaging, findings, differential, diagnosis, dose and management, and matches on specific diseases, signs, drugs and the like; when a question has no specific topic, no sources are forced in. Once an answer is complete, only the sources cited in the text as [1], [2] and so on are listed, with their numbers unchanged; other retrieved entries are folded into “N more entries found, not cited in the answer · View”. Manual search on the knowledge base pages is unchanged.
Yuan Doctor: Q&As added to the knowledge base from the agents now show the full answer. For Q&As submitted with “Add to knowledge base” in the Imaging and Emergency & Critical Care chats, the detail page no longer shows only an opening summary; it shows the whole answer formatted as in the chat, with headings, bold text, lists and tables, and no stray ## or ** marks. Q&As submitted earlier pick this up automatically, with no need to resubmit. Each Emergency & Critical Care answer now also has “Copy answer”, “Discuss in the peer square” and “Add to knowledge base” underneath, and submissions go to review first as before. Q&A detail pages carry a standing note that the entry was generated by a Yuan Doctor AI agent, is for reference by licensed physicians only and does not replace clinical judgement. Q&A references only name entries that have passed review, and Q&As submitted by doctors, edited or not, are always treated as AI-generated and never used as evidence for agent answers.
Yuan Doctor: three in-depth digests in the Emergency & Critical Care knowledge base have been corrected after medical review. Only the sentences the reviewers flagged were changed, and no numbers were touched. PDCD consensus guideline: "first" and "first globally recognised" were removed because the original does not say so; the emergency fluid advice now follows the original, which is to avoid high-dose glucose (> 5% dextrose) infusions and not to use lactate-containing fluids (Ringer's lactate/Hartmann's), rather than to avoid all glucose-containing fluids; the source line, which had been cut off, now gives Keshavan et al., J Inherit Metab Dis 2026, with the funder corrected to SSIEM (MetabERN is not a funder); a stray "。," left from joining text in the summary was removed; and the page date is now the online publication date, 2026-10-02, instead of the 2026-01-01 placeholder. Chinese expert consensus on V/Q scintigraphy for pulmonary thromboembolism: "the first in China" and the sentence saying it does not apply to anticoagulation decisions in confirmed acute PTE ("CTPA preferred") were removed because the abstract does not contain them; the issuing body now reads the Chinese Thoracic Society of the CMA together with several other organisations; and for a positive skin test in patients with protein allergy, "must not be performed" now reads "perfusion imaging is not recommended", as in the original. PARAMEDIC-3 Bayesian secondary analysis of intraosseous versus intravenous access: the phrase saying it does not apply when IV access cannot be obtained was removed because the abstract does not contain it.
Yuan Doctor: the Imaging agent’s report check gets a further set of checks. The typo list is longer; variant spellings such as 影象 and 毛玻璃, together with 脑梗塞, 粘膜, 副鼻窦 and 腔梗, which were already flagged, now all appear as “规范术语建议” (standard-term suggestions) only: they are not replaced in the revised text and do not lower the score. New checks: units and dates; whether the number and unit fit for common measurements such as endometrial thickness, common bile duct width and spleen thickness, and whether a CT value is physically possible; anatomical terms (a right-sided spleen, left/right written for the spinal cord or dural sac, a liver segment that does not match the lobe; no prompt when the report states situs inversus, polysplenia or splenosis); and agreement between findings and impression (spondylolisthesis grade, nodule size wording, a bilateral impression with one-sided findings, “solid” or “stable” in the impression with nothing in the findings, an impression of no abnormality against positive findings). The existing measurement check gives fewer false alarms: lengths over 30 cm such as the extent of a dissection, distance from the incisors or a giant mass are no longer flagged, and measurements that can be under 1 mm, such as carotid intima-media thickness, postmenopausal endometrium or calcifications, no longer trigger the “under 1 mm” prompt. Results show a QC score (excellent / good / fair / poor): a missing impression counts as a major item, a missing measurement or classification as a moderate one, and items raised only by the AI and not by the rules count half; a critical finding in the findings that is missing from the impression gets a separate prompt, and the AI only adds a “please check” note after the original impression rather than writing a diagnosis for you; if an AI edit would put into the impression a critical-finding diagnosis that the original impression does not contain (such as pneumothorax, pulmonary embolism or aortic dissection), whether as an addition or a wording change, it is not written into the revised text: it is only listed as an issue (for a wording change, the suggested wording is shown), and you check and decide. Photo check: a list number at the start of a line that appears on only one side is not counted as a difference; an extra space inside a number (“1 5mm” against “15mm”) is listed so you can check it against the photo. QC does not start while 〔…〕 marks are still in the text box; you are told how many are left to fix. The AI reading of a photo waits up to 55 seconds.
Yuan Doctor: admin review of knowledge entries now leaves a record. Every status change made in Knowledge base management or Agent management (approve, reject, back to pending) logs the time, who made it (an admin doctor account is logged by its account number; a session using only the admin token is logged as “admin token”), the entry number, the status before and after, and a batch number shared by everything done in one action. Actions that were blocked or changed nothing are not logged. Tokens, entry text and rejection reasons are never stored, and records are kept for 365 days. A new “Review log” section at the bottom of each agent’s management page can be filtered by entry number or batch number. If a batch was a mistake, it can be reverted on the server by batch number; only entries nobody has touched since are reverted, and the revert is logged too. No entry content has changed.
Yuan Doctor: admin review of knowledge entries now has a few safeguards. Bulk approve in Knowledge base management and Agent management only acts on pending entries; rejected, already approved and binned entries are skipped, and the result shows how many were approved and how many skipped. The Rejected tab no longer offers bulk approve. A rejected entry can be moved back to approved or pending only one at a time, through a confirmation box that shows the original rejection reason; the server applies the same check. Bulk actions no longer wipe an entry’s existing rejection reason, and a single rejection can carry a short reason (leave it blank to keep the old one). This applies to both the Imaging and the Emergency & Critical Care agents. No entry content has changed.
Yuan Doctor: in the admin Agent management and Knowledge base management lists, tapping the title of a pending or rejected entry no longer opens a not-found page. After the public detail page was closed to unreviewed entries, admin titles still linked to it, and the admin sign-in does not carry over to public pages, so on a phone the tap showed “entry not found”. Tapping such a title now opens an admin preview inside the card: the structured digest section by section, with keywords, source and “Source ↗”, read only through the existing admin API; the summary and key points already on the agent management card are not repeated. Titles of approved entries still open the public detail page. For signed-out visitors and ordinary accounts the public detail page still treats unreviewed entries as not existing, for both the Imaging and the Emergency & Critical Care agents. No entry content has changed.
Yuan Doctor: two changes to “AI 精准识别” (AI precise reading) in the Imaging agent’s report check. First, when reading a photo, misspellings and non-standard wording in the original are now transcribed exactly as written instead of being quietly corrected; corrections appear only under “疑似错别字” (Possible typos) and replace the text only when you tap “采用” (Use). Second, when both an on-device reading and an AI reading are available, the two texts are compared character by character. Each difference is marked in the text box as 〔①本机：…｜AI：…〕 (on-device / AI) and listed below, and for each one you choose “用本机” (use on-device) or “用 AI” (use AI); neither version is applied automatically, and the check will not start while any difference is still open. When the two readings differ too much, choose one whole version against the original image and then review it. Photos and text are still neither stored nor logged, and the check rules have not changed.
Yuan Doctor: entries still under review can now be opened only by the doctor who submitted them (signed in) and by administrators. For the Imaging and the Emergency & Critical Care agents, the detail page, edit page and share count of an unreviewed entry behave for everyone else exactly like an entry that does not exist: the page is not found, and nothing shows whether the entry is there. The briefing and the daily knowledge card on the agent pages now include reviewed entries only, and the briefing’s “new in the last 7 days” counts reviewed entries only. Answers cite reviewed entries only by default. Reviewed entries can be read and shared as before. No entry content has changed.
Yuan Doctor: the Chinese home page now has a “知识库” (Knowledge base) button between Labs and Workbench. The Knowledge base page puts the reviewed entries of the Imaging and the Emergency & Critical Care agents in one place. You can filter by type (guidelines and consensus, literature updates, differential diagnosis, grading and scores, protocols and procedures, drug safety, anatomy and normal values, articles), by agent and by body system, and sort by newest, most read or most shared, 20 entries per page. The second category is “我的专业” (My specialty), the same as in App 1.8.0: entries are ranked by the subspecialty, department and seniority you set under “My preferences” in each agent, together with the topics you ask about most and what you have read in the Knowledge base. Each card states why it is suggested, entries you have opened are marked “已看过” (Read) and moved down, and related entries from the other agent are included. When you type a search term, “为你推荐” (Recommended for you) becomes “相关” (Relevant). Signed in, your preferences follow your account; signed out, they stay in this browser only. Reading history stays on your device. Filters and the page number stay in the address. On phones the layout follows the App; on computers the filters sit on the left and the cards in two columns on the right. No entry content has changed.
Yuan Doctor: photo recognition in the Imaging agent's report check has been upgraded. When AI recognition is enabled on the server, you can tap “AI recognition” right after choosing a photo: if you have framed an area, only that area is uploaded; if not, and the recognition model is already stored on your device, the two boxes “Findings” and “Impression” are located on the device first and only that part is uploaded; otherwise the whole image is uploaded, and the model transcribes only the findings and impression and returns no identifying information. The image is discarded after use and the server neither stores nor logs it. The on-device recognition model is now kept in the browser, so it is downloaded once and never again; browsers that do not support the new recognition engine no longer download the model and go straight to the backup engine. In addition, the “review signature” and “author signature” buttons on the right side of the screen are no longer mistaken for the signature line and no longer cut the report text short. The check rules and critical-value alerts are unchanged.
Yuan Doctor: the Emergency & Critical Care agent no longer cuts off long answers halfway. Complex questions with several coexisting problems or step-by-step management often stopped mid-answer; they now run to the end. With the “detailed teaching” style the answer can be longer, and the time limit has been raised so a long answer is not cut off by a timeout. The sources and calculations behind the answers are unchanged.
Yuan Doctor: the agents page and knowledge search now show reviewed entries only. Knowledge search for the Imaging and the Emergency & Critical Care agents (used by the knowledge lists on the web and in the App) no longer lists entries that are still under review. The entry count and the “new in the last 7 days” figure on the agents page also count reviewed entries only, so both numbers are lower than before. An entry appears as soon as it passes review. No entry content has changed.
Yuan Doctor: in the Imaging and the Emergency & Critical Care agents, a long answer no longer stops halfway when the connection drops. The web page now resumes it once automatically; if that also fails, the answer shows “网络中断（已收到 N 字）” (connection lost, N characters received) with a “继续生成” (continue) button that picks up where the text stopped without repeating what is already shown, instead of the English “network error”. The same button also appears when an answer reaches the single-response length limit. During long pauses in an answer the server now keeps the connection alive, so long answers are less likely to be cut off on mobile networks or VPNs.
Yuan Doctor: the Emergency & Critical Care agent's knowledge base now gives in-depth digests, ordered the way emergency and ICU doctors read: one-line conclusion, key points, population and setting, diagnostic criteria and key thresholds, management steps and drug doses, red flags and must-not-miss conditions, study design and key results, evidence type, grade and limitations, what changes at the bedside, and source guidelines or consensus statements. Sections the source does not cover are marked “not mentioned in the source”, and every number comes from the source. Open-access articles (PMC Open Access, CC-licensed) and LITFL posts include a full Chinese translation with attribution and a link to the original; paywalled articles get a digest of the abstract with a link. When a journal RSS item has no abstract, it is looked up in PubMed and Crossref and clearly labelled if none is found; letters to the editor and masthead pages are no longer digested; English titles are translated into Chinese. Existing entries are being re-digested in the new format.
Yuan Doctor: in the admin, “Agent knowledge” is renamed “Agent management” and opens to one card per agent showing items pending review, approved items, items added in the last 7 days, subscribed sources and the last fetch time; each card leads to that agent’s own management page for fetching now, managing sources, and reviewing and editing items. The Emergency & Critical Care agent now has fetching and review as well: new content from subscribed emergency and critical care journals and professional blogs is fetched automatically and goes to pending review with a structured Chinese digest, and paywalled journals keep only the abstract and the original link. Public pages are unchanged.
Yuan Doctor: share preview images for tool pages are now drawn with fonts bundled on our own server instead of being downloaded from an overseas font service on each request. Previously, when crawlers fetched Korean or Japanese pages, or titles containing ≥ or ≤, a preview could miss characters or fail to generate; Chinese, Japanese, Korean and common math symbols now display correctly. The emoji icon on the left is replaced by the 元 mark and the lock icon in the footer is removed; each image is cached on the server for a day, so repeat fetches are not redrawn. Japanese and Korean tool pages now show the header, tool name and description in their own language, with a short generic line for tools that have no description in that language yet; the French badge is now in French; the tool count in the French and Spanish header is now taken from the catalogue (currently 876) instead of a fixed "480+"; and the stray box around the site address in some previews is gone. Tool content and calculations are unchanged.
Yuan Doctor: when agent knowledge pages are submitted to Baidu for indexing, items that were deleted or not approved are no longer submitted, and links Baidu rejects as invalid or not belonging to this site are marked as failed instead of sitting at the head of the queue. If the daily routine has to resend its report after a network timeout, that report is no longer counted twice against the day's quota for manual submissions in the admin.
Yuan Doctor: agent knowledge items can now be edited by their submitter or an admin (title, summary, key points, tags, source); imaging and emergency knowledge is managed in one admin list, with deletions going to a recoverable trash; the agent knowledge page gains a "Today's picks" block. Fixed a display issue on knowledge detail pages and removed garbled boxes from titles.
Yuan Doctor: launched the Emergency & Critical Care agent (/zhinengti/jizhen) for emergency, ICU, EICU and prehospital physicians. Ask with the case and its numbers, and it runs our clinical calculators to compute qSOFA, SOFA, shock index, P/F, GBS and other scores, shown as cards and used as the basis of the answer; the numbers come from the tools, not from AI estimates. It also has a bedside quick-calc card and an emergency knowledge base. Photo recognition in the imaging report check is improved: you can crop the report body; before recognition the device removes colored stamps and screen moiré; afterwards only the exam, findings and impression are kept, with interface leftovers, character counters, signatures and identifiers removed and common recognition errors fixed. Photos are still recognized on the device only, never uploaded, and recognition does not call a large model. QC rules and the AI review are unchanged; report text is still neither stored nor logged.
Yuan Doctor: the rule check in the imaging report check gains six items; all of them only flag, none changes the text. First, negation reversal: the findings say 'no pneumothorax' but the impression says 'right pneumothorax', or the findings describe a hyperdense temporal lesion but the impression says 'no intracranial haemorrhage'. These are flagged as high risk and are not downgraded when the AI review does not mention them. Second, the exam is listed as non-contrast but the text describes enhancement or arterial, portal venous or delayed phases; sentences citing an outside contrast study or recommending a contrast scan do not count. Third, comparisons: 'larger than before' while the current measurement is smaller than the prior one, or 'smaller' while it grew. A difference of 1 mm or more is flagged as high risk; under 1 mm may be measurement error and is flagged as medium. A prior study dated after the current exam is also flagged. Fourth, two diameters in the findings but a single number in the impression: flagged only when it matches neither the longest diameter (within 0.1 mm) nor the mean diameter (within 0.5 mm). For lung nodules the size category is also checked against the guideline that applies: incidental nodules follow Fleischner, with the mean diameter rounded to the whole millimetre, so 7 mm × 4 mm written as 6 mm is not flagged; low-dose CT screening follows Lung-RADS, with the mean kept to one decimal, so the same 6 mm is really 5.5 mm, crosses a category boundary and is flagged as high risk. Fifth, breast: the impression gives BI-RADS 1 or 2 while the findings mention fine pleomorphic calcifications, spiculation, irregular shape, non-parallel orientation or architectural distortion; this is flagged as 'category may be too low, please check', and no category is assigned. 'No spiculation', 'no architectural distortion', 'regular shape' and distortion at a surgical scar do not count. Sixth, two false alarms are gone: 'cystic low-density lesion in the upper pole of the left kidney' with 'left renal cyst' is no longer flagged as a side mismatch, and scan parameters such as '1 mm reconstruction thickness' are no longer read as lesion size. On CT, a cystic renal lesion without a CT attenuation value is flagged as 'CT value or Bosniak class missing'. Low-dose CT lung cancer screening reports list Lung-RADS only, without Fleischner. Categories and follow-up intervals still come only from our calculators. Report text is still not stored or logged.
Yuan Doctor: the imaging report check now shows the rule check and our calculator results first; the AI review follows, so there is no blank wait. The AI now returns only point-by-point edits, which the site applies after matching them against the original text; an edit that does not match is dropped on its own, and the remaining edits and rule results are still shown. Categories and follow-up intervals still come only from our calculators; any the AI writes in are replaced with a fill-in placeholder. The AI review waits at most 45 seconds; after that it stops, shows an 'AI polishing timed out' notice and keeps the rule results. The rule check also gained three checks: left/right is now compared lesion by lesion, not only for stones (echogenic foci, nodules, masses, space-occupying lesions and cysts count too); breast reports that say only 'left breast', 'right breast' or 'both breasts' are flagged when BI-RADS is missing, without filling in a category; and CT or MR reports that use ultrasound terms such as 'hyperechoic' are flagged as a modality–terminology mismatch, without changing the text. Report text is still not stored or logged.
Yuan Doctor: we now count returning visitors and successful doctor sign-ins per day. Your browser keeps only the date of your last visit (yd_lv), with no identifying information, and you can clear it at any time; the server stores daily totals only, with no account, IP address or device identifier. Bot and monitoring traffic is not counted. The privacy policy now says so.
Yuan Doctor: the retired community-square upload-token endpoint (/api/square/upload-token) now returns 410 Gone; images keep uploading through the site's own upload endpoint.
Visit stats read path: /api/admin/visits accepts excludeBot=1|human=1 (default exclude non-human); write path tags visit_kind (human/bot/ops_ip/rate/day_cap) and keeps rows; daily IP/PV count humans and soft-drop same-day high-freq fake browsers. Still HTTP 200, no access block.
Visit stats: skip more non-human PV (broader UA denylist incl. Playwright/Puppeteer; per-IP day cap default 60 via VISIT_MAX_PV_PER_IP_DAY; quality/overview queries exclude known bot UAs). Still HTTP 200, no access block; no historical row deletes.
- gcs · nihss · sofa
Hot-search lite gate: /api/search/hot again ranks by DB volume; clinical hit-check uses slim clinical-tools-index only (no full site-index load) to avoid better-sqlite3 assert/502; still falls back to GCS/NIHSS/eGFR… when fewer than 6 survivors. Recording uses the same lite gate.
- gcs · nihss · sofa
Homepage hot-search chips fix: /api/search/hot keeps only terms that hit the clinical catalog (tools/scenes/…), otherwise falls back to GCS/NIHSS/eGFR…; recording uses the same gate so patient/consumer terms stop polluting the table; empty-results page no longer links to the closed /search encyclopedia. Run the hot_searches cleanup script once on the server.
- cha2ds2vasc
Site search (/api/search, /zhaogongju and locale search pages) now normalizes query and index text the same way: case, hyphens/spaces/underscores/dots, subscript digits and full-width characters (CHA2DS2-VASc, cha2ds2vasc and CHA₂DS₂-VASc all match); each tool appears once, keeping its best-scored row. No formula or copy changes.
- nihss · china_par
Tool-page verification display: with a structured check, shows 'Verified: date · Available/Degraded/Unavailable/Evidence under review' plus collapsible notes; date-only shows grey 'Result not structured'. First tools: nihss (Available) and china_par (Available after medical recheck). No formula changes.
Fix empty /api/search: use same clinical site-index as /zhaogongju; /api/search-index skips count:0 placeholder JSON and rebuilds at runtime.
ResPic landing/workbench: remove editable SVG/PPTX sell points; share step and FAQ keep PNG + fail-refund only; hide SVG/PPTX primary export buttons (implementation kept).
- nihss · qsofa · cha2ds2vasc
SSR win-search trio: NIHSS / qSOFA / CHA₂DS₂-VASc share a factory for title·H1·can/cannot·doctor FAQ·WebPage+FAQPage JSON-LD (six locales; zh+en full copy).
Short link /r/{key}: tool keys → /linchuang or /{lang}/tools; scene keys still open bedside scenes; unknown → scene list.
Visit stats: ops UA containing YuanDoctor-Ops still returns 200 but skips human PV; optional hot-read VISIT_SKIP_IPS.
Pin dataset review date to 2026-09-20 in source; ignore an older NEXT_PUBLIC_DATASET_UPDATED so llms/changelog.json cannot freeze on a stale date.
Yuan Finance: remove doctor-facing QA mock-credit UI; mock recharge API is admin-only.
Changelog GEO: machine entry points + clinical highlights + structured fields (llms/for-ai link Changelog and changelog.json; clinical section on top; optional tools/DOI/impact).
Impact: Models can extract Changelog URL, dataset review date, and recent clinical entries.
AI-tone cleanup: removed generic AI/marketing wording from the manifest, 404, About, scene, and reproductive-health entry copy while retaining clinical scope, local compute, and usage boundaries.
My Drawer dedupe: stop dual tool+guideline usage on /linchuang that doubled chips; overview favorites/hot/recent deduped.
ResPic Wave A: real SVG (V1 editable title/caption) and real PPTX export; fail→refund; figure light-check + use-statement TXT (no extra charge).
Workbench stickiness: favorites/frequent/recent raised on overview; post-calc next-step chips; tool landings bump recent usage.
App download copy: ZYITONG + Harmony Android-compat path; package com.yuandoctor.app; stores not listed; legacy-brand APK removed from public web.
English/Chinese clinical search 502 hardening (cap 40, SSR try/catch).